LibreEdgeTAM - third-party attribution
======================================

LibreEdgeTAM calls the EdgeTAM implementation provided by the Apache-2.0
Transformers package. LibreYOLO does not vendor the EdgeTAM model source.

    Project:  EdgeTAM
    Upstream: https://github.com/facebookresearch/EdgeTAM
    Commit:   7711e012a30a2402c4eaab637bdb00a521302c91
    License:  Apache License 2.0
    Copyright (c) Meta Platforms, Inc. and affiliates

The upstream project explicitly grants both its code and model checkpoints
under Apache-2.0. The official checkpoint used for the LibreYOLO mirror is:

    Repository: facebook/EdgeTAM
    Revision:   14d7ecc48c656b94e5184519f698cd5386c5a2bf
    File:       edgetam.pt
    SHA-256:    ed2d4850b8792c239689b043c47046ec239b6e808a3d9b6ae676c803fd8780df

``libreyolo/models/sam/edgetam.py`` reproduces the image preprocessing from
``sam2/utils/transforms.py`` at the pinned EdgeTAM commit: square bilinear
resize to 1024 followed by ImageNet normalization, plus the upstream
divide-then-multiply prompt-coordinate scaling. It uses Transformers for the
model implementation and prompt metadata.

The Transformers conversion mapping used by
``weights/convert_edgetam_weights.py`` is adapted from the Apache-2.0
Hugging Face Transformers converter at commit
bd37c453544e83eb875ed3608980a1660376007a. Conversion performs lossless key
remapping plus tensor splitting/concatenation; learned numeric values are not
retrained. The produced state dict is checked tensor-by-tensor against
``yonigozlan/EdgeTAM-hf`` revision
c266ce53b3fc00f0f495b583f6a116c4e57f53bb before publication.

The mirror copies ``.gitattributes`` and the four JSON configuration/processor
files from that same hash-pinned ``yonigozlan/EdgeTAM-hf`` revision. Its model
card declares Apache-2.0. The learned weights are independently converted from
the official checkpoint and then checked exactly against the reference.

The Apache License 2.0 text is included at ``licenses/Apache-2.0.txt``.
